Maryland Citizens Hit With $2B Power Grid Upgrade for Out-of-State AI Data Centers: Who Pays for AI's Energy Crisis?

Published: May 11, 2026 | Updated: May 11, 2026

Update (May 11, 2026): A major controversy is brewing in the energy and tech world. The PJM Interconnection, the regional transmission organization serving 65 million people across 13 states and Washington D.C., has proposed a staggering $2 billion grid upgrade — and Maryland residents are being asked to foot the bill, primarily to support AI data center electricity demand originating outside the state. This story has hit the front page of Hacker News and been widely covered by Tom's Hardware, Ars Technica, and other outlets. Here's everything you need to know.

1. What Happened: The $2B Question

In early May 2026, the PJM Interconnection filed a proposal with the Federal Energy Regulatory Commission (FERC) to approve a massive transmission upgrade across the mid-Atlantic region. The total cost: approximately $2 billion. The justification: accommodating new electricity demand, primarily from AI data centers that are being planned or built in Virginia and other PJM-served states.

The catch? Under PJM's cost allocation rules, Maryland ratepayers — residential and small business customers — would bear a disproportionate share of these costs, even though the data centers that will consume the power are located in other states like Virginia and Pennsylvania.

Maryland Governor Wes Moore's administration has pushed back, arguing that the state's residents should not be forced to subsidize AI infrastructure that primarily benefits Big Tech shareholders. Consumer advocacy groups have called it a "wealth transfer from working families to Silicon Valley."

PJM counters that grid reliability is a shared responsibility across the region, and that the transmission upgrades will benefit all ratepayers through improved grid stability. Critics note that PJM's own analysis shows the primary load drivers are data center interconnection requests, not general reliability needs.

2. How PJM's Capacity Market Enables This

To understand how Maryland residents ended up on the hook for out-of-state AI data centers, you need to understand PJM's unique cost allocation model.

PJM Interconnection operates the largest competitive wholesale electricity market in the world, serving:

PJM uses a regional cost sharing model for large-scale transmission upgrades. When new generation or large loads (like data centers) require transmission system upgrades, the costs are spread across all ratepayers in the region based on formulas set by FERC-approved tariffs — not proportionally to who benefits.

The specific mechanism here is the PJM Regional Transmission Expansion Plan (RTEP). Under RTEP, if PJM determines that a transmission upgrade is needed for reliability, the costs are socialized across the entire region. This worked reasonably well when the upgrades served broadly distributed load growth. But in the age of AI data center concentration, it creates a perverse incentive: states that host data centers get the economic benefits (jobs, tax revenue), while neighboring states get the transmission bills with none of the upside.

A similar controversy erupted in 2024-2025 when PJM's capacity market prices spiked 800% due to data center demand projections, causing rate hikes across the region. The $2B grid upgrade is arguably an even bigger escalation.

3. The Real Cost of AI's Energy Appetite

This controversy isn't happening in a vacuum. AI's electricity consumption is growing at an unprecedented rate, and the numbers are staggering:

Metric Value Source
Single GPT-4 training run energy ~50 GWh Goldman Sachs Research, 2025
Data center electricity (US, 2022→2026) 4% → 9% of US total EIA / Goldman Sachs Research
AI inference per-query energy vs traditional search ~10x-30x more IEA / SemiAnalysis
A single 1 GW AI data center annual consumption ~8.76 TWh Equivalent to ~750K US homes
US data center power demand by 2027 (projected) ~50 GW McKinsey, 2025
PJM queue: data center interconnection requests ~75+ GW PJM IRP backlog, 2026

The PJM interconnection queue is bursting with data center projects — over 75 GW of new requests as of early 2026, a significant portion of which are AI-related. For context, PJM's entire current peak demand is about 165 GW. These data centers would effectively increase the grid's burden by nearly 50%.

AI inference is particularly energy-intensive. A single ChatGPT query uses approximately 10-30x the electricity of a traditional Google search. As AI agents, multimodal models, and real-time inference become ubiquitous, the per-query energy cost compounds dramatically. When you consider that agents may execute hundreds of inference calls per task, the energy footprint scales even faster.

Related reading: Local AI Should Be the Norm: Why Cloud APIs Are Breaking Your App — which covers the energy and cost trade-offs of cloud vs. on-device AI inference.

4. The Environmental Price Tag

The $2 billion figure is just the transmission cost — it doesn't include the generation, cooling, or environmental costs of actually running these data centers.

Carbon emissions: Despite aggressive renewable energy pledges from Big Tech, the reality is more complex. Data centers require 24/7 reliable power, and renewables alone (solar, wind) cannot provide that without massive battery storage or natural gas backup. In PJM's territory, where coal plants are still in the generation mix, every additional GW of data center load risks prolonging the life of fossil fuel plants.

A 2025 study from the Electric Power Research Institute (EPRI) found that data center load growth in PJM could delay coal plant retirements by 3-7 years under some scenarios. This is at odds with Maryland's ambitious climate goals (60% emissions reduction by 2031).

Water consumption: AI data centers use significant amounts of water for cooling. A typical 150 MW data center consumes 3-5 million gallons of water per day. In the mid-Atlantic region, where droughts are becoming more common, this adds another layer of environmental stress.

Land use: The transmission lines required for the $2B upgrade would cut through Maryland communities, requiring new right-of-way corridors. Local opposition to new transmission lines is already mounting in several Maryland counties.

E-waste: AI hardware (GPUs, TPUs, custom ASICs) has a much shorter lifespan than traditional server infrastructure — often 2-3 years before being replaced by newer, more efficient generations. This creates a growing e-waste stream that is rarely factored into AI's environmental cost calculus.

5. Who Should Really Pay for AI Power?

The core ethical question: Should residential electricity ratepayers subsidize AI infrastructure that primarily benefits corporate shareholders?

The case for data center "beneficiary pays":

The case for continued socialization:

What consumer advocates are saying: Organizations like the Maryland Office of People's Counsel and the National Consumer Law Center argue that the current system is fundamentally broken. Residential customers, who have no choice in their electricity provider, are being asked to subsidize the infrastructure needs of multi-trillion-dollar tech companies.

"This is a textbook case of cost-shifting. The companies building these data centers have market capitalizations in the trillions. They can afford to pay for their own grid connections. Instead, they're asking Maryland families on fixed incomes to pick up the tab."

6. The Local AI Alternative: Why On-Device Inference Matters

This controversy raises a deeper question: What if AI didn't need to be this energy-intensive in the first place?

The current AI paradigm is heavily centralized: all inference happens in massive cloud data centers. But a growing movement advocates for on-device AI — running inference locally on consumer hardware rather than in the cloud.

Key advantages of on-device AI:

Tools like Ollama, LM Studio, and Apple's native FoundationModels framework are making on-device AI practical for a growing range of tasks. While large-scale training will always require data center resources, a significant fraction of inference workloads — perhaps 50-70% — could run locally without meaningful quality degradation.

Read our comprehensive guide: Local AI Should Be the Norm: Why Cloud APIs Are Breaking Your App for a full analysis of on-device AI capabilities.

For developers working with AI agents: Ollama & Open WebUI: The Complete 2026 Local AI Stack Guide covers how to set up a fully local AI infrastructure.

7. Timeline of Events

8. What Happens Next

The battle over who pays for AI's grid costs is only beginning. Here's what to watch:

  1. FERC decision: The $2B proposal will ultimately be decided by FERC, which has shown increasing willingness to re-examine cost allocation methodologies. A decision is expected in Q3 2026.
  2. State legislative action: Several states in PJM's footprint — including Maryland, Virginia, and Illinois — are considering legislation to limit cost-shifting to residential ratepayers.
  3. Local AI adoption: As energy costs rise, the economic case for local AI inference strengthens. We may see accelerated investment in on-device AI hardware and software.
  4. Data center regulation: Expect more states to follow the lead of Georgia and South Carolina in requiring data centers to pay their full interconnection costs.

The $2B question isn't just about Maryland. It's about whether the AI industry's infrastructure costs should be socialized or privatized — a debate that will define the energy landscape for the next decade.

Related Articles

First published: May 11, 2026. Sources: Tom's Hardware, PJM Interconnection, FERC filings, Maryland Office of People's Counsel, Goldman Sachs Research, EPRI, IEA.